https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108386
Bug ID: 108386
Summary: Missed optimization with -fno-omit-frame-pointer on
x86
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
void bar (char *);
void
foo (void)
{
char buf[384];
bar (&buf[0]);
bar (&buf[16]);
bar (&buf[127]);
bar (&buf[128]);
bar (&buf[256]);
bar (&buf[380]);
bar (&buf[384]);
}
compiles with -O2 -fno-omit-frame-pointer to:
pushq %rbp
movq %rsp, %rbp
subq $384, %rsp
leaq -384(%rbp), %rdi
call bar
leaq -368(%rbp), %rdi
call bar
leaq -257(%rbp), %rdi
call bar
leaq -256(%rbp), %rdi
call bar
leaq -128(%rbp), %rdi
call bar
leaq -4(%rbp), %rdi
call bar
movq %rbp, %rdi
call bar
leave
ret
but this is unnecessarily large. As frame pointer is here only because user
asked for it, the compiler knows there is always constant difference between
the stack pointer and frame pointer and perhaps in machine reorg could
interchange those cases which would be smaller and not slower.
For these particular leaq/movq instructions, movq %r{sp,bp}, %rdi is 3 bytes,
leaq SIMM8(%rbp), %rdi 4 bytes, leaq SIMM8(%rsp), %rdi 5 bytes, leaq
SIMM32(%rbp), %rdi 7 bytes and leaq SIMM32(%rsp), %rdi 8 bytes.
So at least from code size POV, movq %rsp, %rdi is smaller than any %rbp based
leaq,
and similarly leaq SIMM8(%rbp), %rdi 2 bytes smaller than leaq SIMM32(%rbp),
%rdi.
So, the mov would be a win always, and for frame sizes of more than 128 bytes
%rsp up to 127 offset too.
Though, I think the aliasing code hardcodes frame pointer knowledge and ditto I
think
the unwinder would be quite upset if we did such changes early.